| name | grill-me |
| description | Sharpen a plan or design through self-interrogation when no human reviewer is available. |
| disable-model-invocation | true |
Self-Grill
CI mode: no human is available to interview. The agent runs a self-grill: it generates the design or plan, then plays both the interviewer and the answerer to stress-test it, recording every question and answer inline so the result is auditable.
Process
- State the artifact. Write down the plan, spec, or design being grilled — one paragraph at the top of the response. If the input is a loose idea rather than a written plan, draft the plan first.
- Generate the questions. Produce a numbered list of tough, specific questions a skeptical reviewer would ask: edge cases the plan doesn't cover, terms left undefined, contradictions, missing acceptance criteria, unstated assumptions, scope leaks, prior-art opportunities. Generate at least 10; stop when no new questions surface.
- Answer them honestly. For each question, give the most plausible answer grounded in the codebase, the issue body, and any linked material. If the question genuinely cannot be answered without a human, mark it
needs-human and list the missing input. Do not invent answers; flagging the gap is a valid resolution.
- Apply the answers. For every answered question, update the artifact so it absorbs the answer — tighten the wording, add the missing edge case, define the undefined term, narrow the scope. Show the before/after in the response.
- Surface open items. Conclude with the list of
needs-human questions and any decisions you marked with low confidence. These are the items a human reviewer would still need to confirm.
The output is the hardened plan, plus an audit trail of the questions and answers that produced it.